Speech components in the phonetic characterisation of speakers: a study on complementarity and redundancy of conveyed information

The decomposition of the speech signal into phonetically meaningful units allows the analysis of between- and within-speaker variations. These are components associated with characteristics whose nature relates to the physical, psychological and social aspects of a speaker. In this thesis, we compare perceptual characterisation results with a phonetic analysis and advanced modelling techniques through Convolutional Neural Networks (CNN). Clusterings’ analysis shows that the perceptual results are coherent with those obtained by the CNN and phonetic approaches, which supports the application of these methods in Phonetics. Our results highlight that spectrograms are the most accurate speech representation for speaker identification (96% correct answers on average). Higher formants and harmonics are more important in the characterisation of female voices. Whereas, voice quality characteristics, such as breathiness and hoarseness, play a major role in the characterisation of male speakers. The comparison between Mel Frequency Cepstral Coefficients (MFCC) and classical phonetic measurements is also examined. The MFCC are mainly linked to intensity and f in the characterisation of female speakers, while to the distributions of energy and low level spectral shape for male speakers. Our findings confirm the importance of describing the within-speaker variation for a more complete un- derstanding of between-speakers differences.